Medical image segmentation method based on multi-scale convolution bidirectional Mama

Abstract

The invention relates to the technical field of image processing, in particular to a medical image segmentation method based on multi-scale convolution bidirectional Mama. The method comprises the following steps: constructing a medical image segmentation model comprising an encoder, a decoder and a jump connection module; a medical image segmentation model is trained by using the medical image, local features are extracted through a CNN branch, and a long-range context dependency relationship is captured through a multi-scale bidirectional Mama branch; and updating the parameters of the medical image segmentation model according to the target loss, and performing medical image segmentation through the trained medical image segmentation model, thereby improving the medical image segmentation effect.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a medical image segmentation method based on multi-scale convolutional bidirectional Mamba. Background Technology

[0002] In the field of medical image segmentation, State-Space Models (SSMs) and their variant Mamba offer novel and effective technical solutions. SSMs encode input signals as latent states and evolve them using linear dynamical systems, enabling the modeling of long-range dependencies with linear computational complexity, significantly reducing computational requirements while maintaining powerful modeling performance. Compared to the Transformer architecture, SSMs are more efficient in high-resolution medical image processing, effectively capturing relationships between spatially distant anatomical regions and overcoming the limitations of Convolutional Neural Networks (CNNs) in capturing global contextual dependencies. Mamba, combining selective scanning mechanisms and hardware-aware optimization, achieves linear-time processing of long sequences, integrating the sequence modeling capabilities of Recurrent Neural Networks (RNNs) with the parallelism of Transformers, providing a practical and scalable solution for long-range dependency modeling in high-resolution imaging tasks. Researchers have also proposed several architectures that integrate SSMs. For example, U-Mamba combines the global modeling capabilities of SSMs with the local feature extraction capabilities of CNNs in a U-shaped architecture, VMamba enhances spatial dependency modeling through a two-dimensional state space mechanism, and VM-Unet constructs a complete encoder-decoder framework based on the principles of SSMs.

[0003] However, existing SSMs-based technologies have key limitations that hinder their practical clinical application in medical image segmentation. First, the traditional Mamba model employs a unidirectional scanning strategy, leading to gradual information decay when modeling long-range dependencies, significantly limiting its ability to comprehensively capture complex spatial relationships in medical images. Second, existing technologies struggle to balance fine-grained local feature extraction with global contextual information modeling, lacking robust multi-scale feature fusion mechanisms, which reduces segmentation accuracy, especially when dealing with anatomical structures and pathological regions with complex morphological variations. Third, most Mamba-based architectures primarily focus on spatial domain features, paying insufficient attention to the key texture and boundary information encoded in the frequency domain. Summary of the Invention

[0004] Based on this, the present invention provides a medical image segmentation method based on multi-scale convolutional bidirectional Mamba, which solves at least one problem in the prior art.

[0005] In a first aspect, the present invention provides a method for training a medical image segmentation model based on multi-scale convolutional bidirectional Mamba, which includes the following steps: A medical image segmentation model is constructed, comprising an encoder, a decoder, and a skip connection module; wherein the decoder includes a multi-scale convolutional bidirectional Mamba module, and the multi-scale convolutional bidirectional Mamba module includes a CNN branch and a multi-scale bidirectional Mamba branch; The medical image segmentation model is trained using medical images, where local features are extracted through CNN branches and long-range contextual dependencies are captured through multi-scale bidirectional Mamba branches. The parameters of the medical image segmentation model are updated based on the target loss.

[0006] In some optional embodiments, the CNN branch includes at least two convolutional layers, and each convolutional layer is followed by a batch normalization layer and an activation layer in sequence; features are sequentially convolved by the convolutional layers, batch normalized by the batch normalization layers, and non-linearly activated by the activation layers.

[0007] In some optional embodiments, the multi-scale bidirectional Mamba branch includes a first normalization layer, a bidirectional Mamba module, a second normalization layer, and a multi-scale feature fusion module; features are sequentially normalized through the first normalization layer, feature representations are learned through the bidirectional Mamba module, residual connections are performed, then sequentially normalized through the second normalization layer, feature fusion is performed through the multi-scale feature fusion module, and finally residual connections are performed.

[0008] In some alternative embodiments, the operation in the bidirectional Mamba module is as follows: ; ; ; ; ; in, This indicates the input characteristics of a bidirectional Mamba module. Representation layer normalization, Indicates linear projection. Indicates SiLU function activation; This represents the processing procedure for the forward state-space model (SSM). This indicates an operation that reverses the order of elements in a sequence. Indicates the backward convolution operation. This describes the processing steps of the inverse state-space model (SSM). This represents the element-wise multiplication operation of matrices. This indicates the output characteristics of the bidirectional Mamba module.

[0009] In some optional embodiments, the operations in the multi-scale feature fusion module are as follows: ; ; ; in, This represents the input features of the multi-scale feature fusion module. This represents a 3×3 convolution with a hole ratio of 2. This represents a 3×3 convolution with a hole rate of 3. Represents a 1×1 convolution. This represents a 3D convolution with a kernel size of 3×3×3. This indicates activation by the nonlinear activation function (GeLU). This represents the output features of the multi-scale feature fusion module.

[0010] In some alternative embodiments, the skip connection module includes a multi-frequency attention (MFA) module, wherein operations within the multi-frequency attention (MFA) module include: Features from the encoder and decoder features By performing fusion, fusion characteristics are obtained. ; Fusion features Adaptive pooling and discrete cosine transform (DCT) are performed to obtain multiple frequency components; For each frequency component, global average pooling, global max pooling, and global min pooling are performed. The components are then combined by adding them element by element and divided by the number of frequency components to generate aggregated features. All aggregated features are aggregated, then processed through a fully connected layer and activated by an activation function to obtain the channel attention map. ; According to the channel attention map Fuse features from the encoder and decoder features Generate feature maps ; Processing feature maps using spatial attention mechanism The output features of the Multi-Frequency Attention (MFA) module are obtained.

[0011] In some optional embodiments, the target loss is calculated by the following formula: ; in, Indicates target loss; and These represent hyperparameters; Represents the Dice loss function. This represents the cross-entropy loss function.

[0012] Secondly, the present invention provides a medical image segmentation method based on multi-scale convolutional bidirectional Mamba, which includes the following steps: The medical image is input into the medical image segmentation model, and the medical image segmentation model outputs the segmentation result; The medical image segmentation model is obtained by the medical image segmentation model training method based on multi-scale convolutional bidirectional Mamba as described above.

[0013] Thirdly, the present invention provides a medical image segmentation device, comprising: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that, when executed by at least one processor, implement the medical image segmentation method based on multi-scale convolutional bidirectional Mamba as described above.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the medical image segmentation method based on multi-scale convolutional bidirectional Mamba as described above.

[0015] Due to the adoption of the above technical solutions, the embodiments of the present invention have at least the following beneficial effects: (1) The medical image segmentation model integrates CNN branches, multi-scale bidirectional Mamba branches and multi-scale feature fusion modules, which breaks through the key limitations of existing methods in local feature extraction, global context modeling and multi-scale feature integration. (2) The multi-scale convolutional bidirectional Mamba module integrates local detail preservation and global semantic understanding functions. Through collaborative local-global feature extraction, it achieves special multi-scale aggregation of complex anatomical structures. (3) The Multi-Frequency Attention (MFA) module utilizes frequency domain representation through Discrete Cosine Transform (DCT) and combines it with spatial attention to improve the feature fusion of encoder and decoder, thereby significantly enhancing the model’s attention to key anatomical and diagnostic regions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the network structure of a medical image segmentation model in one embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the network structure of a multi-scale convolutional bidirectional Mamba module in one embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the network structure of a bidirectional Mamba module in one embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the network structure of a multi-scale feature fusion module in one embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of the network structure of the multi-frequency attention module in one embodiment of the present invention.

[0021] Figure 6 This is a partial visualization of the segmentation results of the MCBM-MFA model and existing medical image segmentation models on the ACDC dataset in one embodiment of the present invention.

[0022] Figure 7 This is a partial visualization of the segmentation results of the MCBM-MFA model and existing medical image segmentation models on the ISIC2018 dataset in one embodiment of the present invention.

[0023] Figure 8 This is a partial visualization of the segmentation results of the MCBM-MFA model and existing medical image segmentation models on the GlaS dataset in one embodiment of the present invention.

[0024] Figure 9 The images show partial visualizations of the MSFF module's segmentation results on the ACDC, GlaS, and ISIC2018 datasets, achieved using 3D and 2D convolutions respectively, in this embodiment of the invention. Detailed Implementation

[0025] The following will provide a clear and complete description of the concept and technical effects of the present invention, so as to fully explain the purpose, solution and effects of the present invention.

[0026] Figure 1 The network structure of a medical image segmentation model according to an embodiment of the present invention is shown. For ease of representation, the medical image segmentation model is abbreviated as MCBM-MFA model.

[0027] Specifically, the MCBM-MFA model includes an encoder, skip connection modules, and a decoder. The encoder can employ a pre-trained PVTv2 (Pyramid Vision Transformer v2) backbone network, consisting of four layers with progressively increasing channel dimensions [64, 128, 320, 512]. Through stacked convolutions and stride operations, the encoder achieves efficient multi-scale feature extraction, while each stage halves the spatial resolution, resulting in feature maps with input resolutions of [1 / 4, 1 / 8, 1 / 16, 1 / 32].

[0028] The decoder comprises upsampling layers and multi-scale convolutional bidirectional Mamba (MCBM) modules. Each MCBM module contains two parallel branches: a CNN-based branch (CNN branch) for local detail feature extraction, and a multi-scale bidirectional Mamba branch for capturing long-range contextual dependencies. The decoder consists of four stages, with the channel dimensions decreasing sequentially [512, 320, 128, 64] to progressively restore spatial resolution. In each decoding stage, features from the encoder and skip connection modules are processed independently by the two branches and then fused, thereby combining fine-grained local features with high-level semantic context. This collaborative design enhances semantic consistency across multiple scales.

[0029] To further enhance the interaction between the encoder and decoder, a Multi-Frequency Attention (MFA) module is introduced as an enhanced skip connection mechanism. In other words, the skip connection module includes the MFA module. The MFA module transforms spatial features to the frequency domain using Discrete Cosine Transform (DCT) and combines it with the spatial attention mechanism to achieve more effective feature alignment and texture preservation. This design significantly improves segmentation accuracy and boundary delineation while maintaining computational efficiency.

[0030] like Figure 2 As shown, the MCBM module includes a CNN branch and a multi-scale bidirectional Mamba branch, synergistically integrating the local modeling advantages of CNNs with the global dependency modeling capabilities of state-space models. The CNN branch employs a two-layer 3×3 convolutional architecture, with each convolutional layer followed by a batch normalization (BN) layer and an activation layer. The activation layer can be activated using the ReLU function. This design progressively extracts hierarchical local features, effectively capturing low-level texture patterns and high-level edge contours while preserving morphological structure. Stacked convolutional layers with non-linear activation generate multi-scale receptive fields, producing fine-grained visual representations, which is particularly crucial for accurate medical image interpretation and analysis.

[0031] The multi-scale bidirectional Mamba branch includes a bidirectional Mamba (BM) module and a multi-scale feature fusion (MSFF) module, both preceded by a layer normalization layer (LayerNorm). The BM module learns sequence feature representations, significantly enhancing the model's ability to capture long-range dependencies and global contextual information. Simultaneously, the MSFF module aggregates features from multiple receptive field scales, achieving robust semantic integration and improving adaptability to different target sizes and complex anatomical structures.

[0032] The processing procedure of the MCBM module in stage i is shown in the following equation: ; ; ; ; in, This represents the input features of the i-th stage MCBM module. Since the decoder contains four stages, i can be 1, 2, 3, or 4. This indicates that 3×3 convolution, batch normalization, and ReLU activation are performed sequentially. , , These represent the bidirectional Mamba module processing, layer normalization, and multi-scale feature fusion module processing, respectively. This represents the output characteristics of the i-th stage MCBM module.

[0033] like Figure 3 As shown, the bidirectional Mamba module employs a two-branch architecture. The first branch performs a linear projection of the input features, followed by non-linear activation using the SiLU function, as shown in the following equation: ; in, This indicates the input characteristics of a bidirectional Mamba module. H, W, and C represent the height, width, and number of channels of the feature map, respectively. Representation layer normalization; Indicates linear projection; This indicates that the SiLU function is activated.

[0034] The second branch first obtains the sequence representation through linear projection. Subsequently, the sequence is processed by the forward and backward state-space model (SSM) modules, as shown in the following equations: ; ; ; in, This represents the processing procedure for the forward state-space model (SSM). This indicates an operation that reverses the order of elements in a sequence. Indicates the backward convolution operation. This describes the processing procedure for the inverse state-space model (SSM).

[0035] Finally, the output of the bidirectional Mamba module is calculated using the following formula: ; in, This represents the element-wise multiplication operation of matrices. This indicates the output characteristics of the bidirectional Mamba module.

[0036] Bidirectional Mamba modules model both forward and reverse contextual dependencies simultaneously, enabling the learning of richer and more comprehensive semantic representations. This bidirectional scanning mode significantly enhances the model's ability to learn spatially sensitive features, achieving higher representation fidelity compared to unidirectional models.

[0037] like Figure 4 As shown, the MSFF module employs a dual-branch architecture. The main branch first expands the channel dimension through 1×1 convolutional layers, followed by a series of multi-scale dilated convolutions to achieve comprehensive feature extraction. These dilated convolutions use a consistent 3×3 kernel size but have different dilation rates to extract fine-grained local features at different spatial resolutions. For the input features... The operation of the first branch is shown in the following formula: ; in, and These represent 3×3 convolutions with hole ratios of 2 and 3, respectively. This represents a standard 1×1 convolution.

[0038] In the second branch, 3D convolutional layers are used to capture and model contextual information and spatial structure, followed by activation using a non-linear activation function (GeLU), as shown in the following equation: ; in, This represents a 3D convolution operation with a kernel size of 3×3×3.

[0039] To obtain the final output of the MSFF module, the features of the two branches are fused by element-wise multiplication, as shown in the following equation: ; in, This indicates the output characteristics of the MSFF module.

[0040] The MSFF module aggregates features from multiple receptive field scales to extract complex multi-scale contextual dependencies commonly found in medical images, enhancing the network's nonlinear expressive power and spatial modeling capabilities.

[0041] like Figure 5As shown, the MFA module first applies Discrete Cosine Transform (DCT) to the fused features of the encoder and decoder. This frequency domain decomposition enables more effective modeling of multi-scale anatomical structures, covering a range from small pathological lesions to large organ systems. and The output features of the encoder and decoder are represented separately. First, they are concatenated to obtain the fused features. Then, it is adjusted to a fixed size using adaptive pooling. As shown in the following formula: ; in, Indicates splicing, This indicates adaptive pooling.

[0042] Next, a set of predefined DCT filters are used to extract frequency features, with each filter corresponding to a specific coordinate in the frequency domain. As shown in the following formula: ; in, u and v These represent the horizontal and vertical frequency components in the frequency domain coordinates, respectively; Indicates the first j A DCT filter is defined as follows: ; ; ; in, and Represents the coordinate position in the spatial domain.

[0043] For each frequency component, global average pooling, global max pooling, and global min pooling operations are applied. Then, all the resulting pooled features are combined by element-wise addition and divided by the number of frequency branches to generate aggregated features. , and As shown in the following formula: ; ; ; in, n Indicates the number of frequency branches. This indicates a global average pooling operation. This represents the global max pooling operation. This represents the global minimum pooling operation.

[0044] Next, , and The aggregated data is then fed into a fully connected layer and processed by an activation function to generate a channel attention map M, as shown in the following equation: ; ; in, , and These represent the activation of the Sigmoid function, 1×1 convolution, and ReLU function, respectively.

[0045] The channel attention map M is then combined with the encoder output features to generate a feature map. As shown in the following formula: ; ; Here, α represents the adaptive adjustment factor. The adaptive adjustment factor is a learnable parameter that dynamically adjusts the channel attention distribution, thereby achieving adaptive and context-aware feature fusion, achieving optimal balance among different frequency feature components, and enhancing the model's adaptability in various types of medical image segmentation tasks.

[0046] Subsequently, to further address the imbalance of spatial information after frequency domain fusion and enhance sensitivity to key anatomical regions, the MFA module integrates a spatial attention module. This spatial attention module improves the contour depiction of lesion boundaries and other salient structures by adaptively reweighting spatial features. Specifically, it first computes max pooling and average pooling in parallel along the channel dimension to extract complementary spatial information. Then, by applying 7×7 convolutions to the stitched features, the model can encode fine-grained spatial interactions. Next, a sigmoid activation function generates normalized attention weights in the range [0,1]. The resulting attention map highlights regions with spatial information, guiding the model to focus on anatomically relevant areas, thereby improving segmentation performance. Spatial attention weights The calculation is shown in the following formula: ; in, This represents the input to the spatial attention module. This represents the Sigmoid activation function. Represents a 7×7 convolution. This represents max pooling along the channel dimension. This represents average pooling along the channel dimension.

[0047] Finally, spatial attention weights respectively with feature map and encoder output features They are multiplied to generate their corresponding weighted representations. These two attention-weighted features are fused by element-wise addition, as shown in the following equation: in, This indicates the output characteristics of the MFA module.

[0048] The parameters of the medical image segmentation model are updated based on the composite loss (target loss). The formal definition of the composite loss is as follows: ; in, Indicates compound loss; and These represent hyperparameters; Represents the Dice loss function. This represents the cross-entropy loss function.

[0049] To verify the effectiveness of the medical image segmentation model in this embodiment of the invention, experiments were conducted using the GlaS dataset, the ISIC2018 dataset, and the ACDC dataset, and the model was compared with several existing medical image segmentation models.

[0050] The Glandular Segmentation (GlaS) dataset is a widely recognized benchmark dataset in the field of colonic gland segmentation, originating from the 2015 MICCAI Glandular Segmentation Challenge. This dataset contains 165 hematoxylin-eosin (H&E) stained microscopic images of colonic tissue sections, covering both benign and malignant samples. Each image has been pixel-by-pixel annotated by professional pathologists to delineate the glandular structure. To ensure consistency between training and evaluation, all images were resized to 224×224 pixels and divided into 85 training samples and 80 test samples. A five-fold cross-validation scheme was used to comprehensively evaluate segmentation performance and generalization ability.

[0051] The ISIC2018 dataset, compiled by the International Skin Imaging Collaboration, is a benchmark dataset for skin lesion segmentation and melanoma detection. This dataset contains 2594 images, encompassing clinical photographs and dermoscopic images, with lesion masks annotated by dermatologists. To ensure consistency, each image was resized to 224×224 pixels. The dataset is divided into training, validation, and test sets in a 70 / 10 / 20 ratio. Experiments employ five-fold cross-validation to comprehensively evaluate the model's generalization ability.

[0052] The ACDC dataset, released at MICCAI 2017, provides comprehensive cardiac MRI annotated images for segmentation tasks. This dataset contains short-axis cine MRI sequences from 100 subjects, including healthy individuals and patients with right ventricular (RV), left ventricular (LV), and myocardial (Myo) disease. To normalize the input dimensions, all slices were resized to 224×224 pixels. The dataset consists of 70 training samples, 10 validation samples, and 20 test samples.

[0053] Image segmentation quality was evaluated using key metrics including Dice coefficient, mIoU, and HD95. The Dice coefficient quantifies the similarity between the segmentation result and the corresponding ground truth annotations. mIoU (mean intersection-union ratio) measures the average overlap between the predicted segmentation mask and the ground truth mask. Both the Dice coefficient and mIoU values ​​range from 0 to 1; higher values ​​indicate a closer similarity to the ground truth annotations, implying higher accuracy and reliability of the model. The 95% Hausdorff distance (HD95) quantifies the maximum boundary difference between the predicted segmentation and the ground truth annotations; lower values ​​indicate a better boundary fit. The specific calculations for these metrics are shown in the following formulas: ; ; ; Where TP represents the number of pixels correctly predicted as the target region, FP represents the number of pixels incorrectly predicted as the background, and FN represents the number of pixels incorrectly predicted as the background. y Represents the actual value. Indicates the predicted segmentation result. and represents the one-way Hausdorff distance from the true value to the predicted result and from the predicted segmentation result to the true value, respectively.

[0054] The experiments were implemented using the PyTorch framework, and all training and testing were performed on an NVIDIA RTX4080 GPU equipped with 16GB of video memory. As shown in Table 1, the input images were normalized to a resolution of 224×224. For optimization, the AdamW algorithm was used with an initial learning rate of 0.001 to accelerate convergence and improve generalization performance.

[0055] Table 1. Hyperparameters used in the experiment Table 2 shows the Dice coefficient results of the MCBM-MFA model and existing medical image segmentation models on the ACDC dataset in the embodiments of the present invention. The Dice scores of the MCBM-MFA model on the right ventricle (RV), myocardium (Myo), and left ventricle (LV) are 89.96%, 89.58%, and 95.99%, respectively, with an average Dice score of 91.83%. Compared with high-performance Transformer-based models (such as PVT-CASCADE and CSWin-UNet, both reaching 91.46%), the MCBM-MFA model achieves a relative improvement of 0.37%, indicating its stronger ability to accurately segment cardiac substructures with diverse morphologies and scales. In addition, the MCBM-MFA model outperforms several hybrid CNN-Transformer architectures, including TransUNet (89.71%) and MISSFormer (90.86%), highlighting the effectiveness of the proposed bidirectional Mamba module and multi-scale feature fusion module in jointly capturing local texture and global contextual dependencies. Furthermore, although pure Transformer-based models (such as Swin-UNet, CSWin-UNet, and PVT-CASCADE) generally outperform traditional CNNs due to their superior global receptive field, they still have limitations in modeling fine-grained anatomical details, especially in regions with blurred boundaries or low contrast. In contrast, the MCBM-MFA model effectively bridges this gap by integrating frequency domain attention mechanisms and spatial context modulation, achieving accurate boundary segmentation and better identifying subtle structural changes. These results collectively demonstrate the robustness and generalization ability of the MCBM-MFA model in cardiac MRI segmentation tasks.

[0056] Table 2. Dice coefficient results of various medical image segmentation models on the ACDC dataset. Figure 6 The qualitative visualization results of the MCBM-MFA model and existing medical image segmentation models in this embodiment of the invention are presented on the ACDC dataset. It can be seen that the MCBM-MFA model has superior capabilities in handling complex cardiac anatomy. Compared with representative models including R50+UNet, R50+AttenUNet, SwinUNet, MISSFormer, and VM-UNet, the MCBM-MFA model achieves significant improvements in both segmentation accuracy and structural fidelity, especially in delineating the boundaries of the right ventricle (RV), myocardium (Myo), and left ventricle (LV).

[0057] The evaluation on the ISIC2018 dataset employed five-fold cross-validation, and the results are summarized in Table 3. It can be seen that the MCBM-MFA model exhibits strong and stable performance in skin lesion segmentation, outperforming many existing medical image segmentation models in both mIoU and Dice metrics. Specifically, in terms of mIoU, the MCBM-MFA model achieved a score of 84.95%, a 0.49% improvement over CFATransUnet; compared to the classic U-Net, the improvement reached 3.72%, indicating that the MCBM-MFA model is significantly effective in accurately delineating lesion boundaries and reducing background interference. In terms of Dice, the MCBM-MFA model ranked first with a score of 91.08%, significantly outperforming the second-ranked CFATransUnet; compared to I2U-Net and Swin-Unet, the MCBM-MFA model achieved performance improvements of 0.98% and 1.47%, respectively, demonstrating superior segmentation reliability and consistency when facing diverse lesion morphologies and contrast variations.

[0058] Table 3. mIoU and Dice coefficient results for each model on the ISIC2018 dataset. Figure 7 This paper presents the qualitative segmentation results of various models on the ISIC2018 skin lesion dataset, highlighting the superior performance of the MCBM-MFA model compared to models such as U-Net, UNet++, MFCNet, and HSH-UNet. In five representative cases with different lesion complexities, the MCBM-MFA model consistently demonstrates significant advantages in boundary accuracy and morphological fidelity.

[0059] The comprehensive evaluation results of each model on the GlaS dataset are shown in Table 4. These results demonstrate that the MCBM-MFA model outperforms many existing models based on CNN, Transformer, and hybrid architectures in segmentation performance. Specifically, the MCBM-MFA model achieves the highest Dice coefficient (91.50%) and the lowest HD95 (6.08 mm). Compared to the high-performance MMUNet, the MCBM-MFA model improves the Dice coefficient by 0.33% and reduces the HD95 by 0.26 mm, confirming its enhanced ability to accurately delineate glandular structures in complex histopathological images.

[0060] Table 4. HD95 and Dice coefficient results for each model on the GlaS dataset. Figure 8The qualitative results of each model on the GlaS dataset are presented, highlighting the superior performance of the MCBM-MFA model compared to several existing medical image segmentation models, especially in handling complex pathological tissue structures.

[0061] To investigate the individual effects of each component in the MCBM-MFA model, ablation experiments were conducted on the ACDC and GlaS datasets. Specifically, six model variants were designed to systematically evaluate the individual and combined effects of the MFA and MCBM modules. Furthermore, the MCBM module was split into two independent components (MSFF module and bidirectional Mamba module) to evaluate their respective effects. Based on the MCBM-MFA model, six variant models were obtained by selectively removing, replacing, or modifying modules, which were used to systematically examine the impact of each module. Model I: The MSFF module was replaced with a traditional MLP module to test the effectiveness of the multi-scale feature fusion mechanism; Model II: Replace the bidirectional Mamba module with a unidirectional Mamba module to evaluate the impact on bidirectional modeling capabilities; Model III: Replace the MFA module with Simple Feature Addition (Add) to evaluate the effect of multi-level feature aggregation; Model IV: The MFA and MSFF modules were replaced with the Simple Feature Addition (Add) and MLP modules, respectively, to verify the joint effect of feature aggregation and multi-scale fusion; Model V: The MFA and bidirectional Mamba modules were replaced with the simple feature addition module and the one-way Mamba module, respectively, to test the combined impact on feature aggregation and sequence modeling; Model VI: Replace the MSFF and bidirectional Mamba modules with MLP and unidirectional Mamba modules respectively to evaluate the joint effect of multi-scale feature fusion and bidirectional sequence modeling.

[0062] As shown in Table 5, on the ACDC dataset, the complete MCBM-MFA model achieved state-of-the-art performance across all evaluation metrics, with an overall Dice coefficient of 91.83%. Specific scores for each chamber were: right ventricle (RV) 89.96%, myocardium (Myo) 89.58%, and left ventricle (LV) 95.99%. Compared to the six variant models, the complete model improved Dice scores by 0.16%, 0.46%, 0.39%, 0.23%, 0.15%, and 0.86%, respectively. Similarly, evaluation on the GlaS dataset further confirmed the robustness and effectiveness of the proposed architecture; the MCBM-MFA model achieved the highest Dice coefficient of 91.50% and the lowest HD95 of 6.08 mm, exceeding the corresponding variant models by 0.36%, 0.54%, 0.25%, 1.14%, 0.78%, and 0.72%, respectively. The consistent quantitative improvements on both datasets demonstrate the indispensability and synergistic effect of the MFA, MSFF, and bidirectional Mamba modules.

[0063] Table 5 Ablation Experiment Results To investigate the impact of different encoders on the performance of the MCBM-MFA model, comprehensive experiments were conducted on the GlaS and ISIC2018 datasets using five existing pre-trained encoders. The results are summarized in Table 6. PVTv2 achieved the highest performance, achieving 91.50% Dice coefficients and 6.08mm HD95 on the GlaS dataset, and 91.08% Dice coefficients and 84.95% mIoU on the ISIC2018 dataset. PVTv2 significantly reduces computational complexity while efficiently preserving multi-scale feature extraction, enabling better segmentation of multi-scale targets in medical imaging tasks and achieving accurate boundary localization.

[0064] Table 6. Impact of different encoders on MCBM-MFA performance The performance differences between 3D and 2D convolutions in the MSFF module were evaluated on the ACDC, GlaS, and ISIC2018 datasets. See Table 7 and... Figure 9 As shown, 3D convolution consistently outperforms 2D convolution across all evaluation metrics, indicating that the 3D convolution architecture can better capture multidimensional spatial context, improving the segmentation accuracy of volumetric medical imaging tasks and predominantly two-dimensional medical imaging tasks.

[0065] Table 7. Experimental results comparing 2D convolution and 3D convolution in the MSSF module. To evaluate the computational efficiency of the MCBM-MFA model, the trainable parameters and floating-point operations per second (FLOPs) of several representative medical image segmentation models were compared. Detailed results are shown in Table 8. The MCBM-MFA model uses 48.62M parameters and requires 10.96 GFLOPs. Although not the lightest model, it effectively balances efficiency with powerful representation capabilities.

[0066] Table 8. Trainable parameters and floating-point operations per second for each model The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention by the same or equivalent means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.

Claims

1. A training method for a medical image segmentation model based on multi-scale convolutional bidirectional Mamba, characterized in that, Includes the following steps: A medical image segmentation model is constructed, comprising an encoder, a decoder, and a skip connection module; wherein the decoder includes a multi-scale convolutional bidirectional Mamba module, and the multi-scale convolutional bidirectional Mamba module includes a CNN branch and a multi-scale bidirectional Mamba branch; The medical image segmentation model is trained using medical images, where local features are extracted through CNN branches and long-range contextual dependencies are captured through multi-scale bidirectional Mamba branches. The parameters of the medical image segmentation model are updated based on the target loss.

2. The method according to claim 1, characterized in that, The CNN branch includes at least two convolutional layers, and each convolutional layer is followed by a batch normalization layer and an activation layer in sequence; features are sequentially convolved by the convolutional layer, batch normalized by the batch normalization layer, and non-linearly activated by the activation layer.

3. The method according to claim 1, characterized in that, The multi-scale bidirectional Mamba branch includes a first-level normalization layer, a bidirectional Mamba module, a second-level normalization layer, and a multi-scale feature fusion module; The features are sequentially normalized through the first normalization layer, learned through the bidirectional Mamba module, and then connected by residuals. They are then normalized through the second normalization layer, fused through the multi-scale feature fusion module, and finally connected by residuals.

4. The method according to claim 3, characterized in that, The operations in a bidirectional Mamba module are shown in the following formula: ； ； ； ； ； in, This indicates the input characteristics of a bidirectional Mamba module. Representation layer normalization, Indicates linear projection. Indicates SiLU function activation; This represents the processing procedure for the forward state-space model (SSM). This indicates an operation that reverses the order of elements in a sequence. This represents the processing procedure of the inverse state-space model; This represents the element-wise multiplication operation of matrices. This indicates the output characteristics of the bidirectional Mamba module.

5. The method according to claim 3, characterized in that, The operations in the multi-scale feature fusion module are shown in the following equation: ； ； ； in, This represents the input features of the multi-scale feature fusion module. This represents a 3×3 convolution with a hole ratio of 2. This represents a 3×3 convolution with a hole rate of 3. Represents a 1×1 convolution. This represents a 3D convolution with a kernel size of 3×3×3. This indicates activation by a non-linear activation function. This represents the output features of the multi-scale feature fusion module.

6. The method according to claim 1, characterized in that, The skip connection module includes a multi-frequency attention module, wherein the operations in the multi-frequency attention module include: Features from the encoder and decoder features By performing fusion, fusion characteristics are obtained. ; Fusion features Adaptive pooling and discrete cosine transform are performed to obtain multiple frequency components; For each frequency component, global average pooling, global max pooling, and global min pooling are performed. The components are then combined by adding them element by element and divided by the number of frequency components to generate aggregated features. All aggregated features are aggregated, then processed through a fully connected layer and activated by an activation function to obtain the channel attention map. ; According to the channel attention map Fuse features from the encoder and decoder features Generate feature maps ; Processing feature maps using spatial attention mechanism The output features of the multi-frequency attention module are obtained.

7. The method according to claim 1, characterized in that, The target loss is calculated by the following formula: ； in, Indicates target loss; and These represent hyperparameters; Represents the Dice loss function. This represents the cross-entropy loss function.

8. A medical image segmentation method based on multi-scale convolutional bidirectional Mamba, characterized in that, Includes the following steps: The medical image is input into the medical image segmentation model, and the medical image segmentation model outputs the segmentation result; The medical image segmentation model is obtained by the medical image segmentation model training method based on multi-scale convolutional bidirectional Mamba as described in any one of claims 1-7.

9. A medical image segmentation device, characterized in that, include: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that, when executed by at least one processor, implement the medical image segmentation method based on multi-scale convolutional bidirectional Mamba as described in claim 8.

10. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed by a processor, implement the medical image segmentation method based on multi-scale convolutional bidirectional Mamba as described in claim 8.

Images